Source-linked AI summary
Explainable and Trustworthy AI for Anti-Money Laundering: A Graph-based Hybrid Framework for Real-World Financial Crime Detection
Ali Shahbazi, Faraz Sasani, Arshia Hossein zadeh, Sheyda safaeimoradi, Hossein Najafzadeh
TL;DR
Rule-based AML monitoring struggles with false positives and laundering patterns that span connected accounts, while prior graph AML work has limited evaluation of explanation faithfulness and calibrated uncertainty. This study evaluates a graph-based detection framework with typology-based explanations and conformal prediction; its IMPROVED model achieves the strongest reported detection results among evaluated models.
Problem
Rule-based AML monitoring produces many false-positive alerts, and prior graph AML studies rarely assess explanation faithfulness against documented typologies or statistically calibrated prediction confidence.
Source: §1 · pp. 3–5Method
The framework compares graph architectures and tabular baselines on a temporally partitioned HI-Small graph, then evaluates explanations with GNNExplainer and uncertainty with Mondrian conformal prediction.
Source: §1 · p. 5, §2.7.1 · p. 20Results
Test AUPRC of 64.85% and Best F1 of 68.42% made IMPROVED the strongest detection model among those evaluated.
Source: §4.1 · p. 34Takeaways & Limitations
GNNExplainer recovered documented typologies with higher fidelity than attention-weight and random baselines, while conformal prediction provides ambiguous sets for human-review triage.
Source: §4.5 · p. 38, §5 · p. 41The evaluation uses a synthetic HI-Small benchmark, so validation on proprietary real-world data is needed to clarify transferability across typologies and noise conditions.
Source: §4.7 · p. 40
Abstract
from arXiv · showhide
Purpose: Money laundering threatens financial systems, while rule-based monitoring suffers from high false-positive rates and limited ability to capture relational transaction patterns. This study proposes an explainable graph-based framework for anti-money laundering (AML) detection that jointly addresses predictive performance, explanation faithfulness, and uncertainty calibration. Methods: Using the IBM Transactions for Anti-Money Laundering (HI-Small) benchmark, comprising 5,078,345 transactions among 518,573 accounts with 0.10% illicit transactions, a directed attributed graph with 4,487,133 edges was constructed using temporal, leakage-safe partitioning. Three GATv2 architectures, BASE, BASE-Large, and IMPROVED, incorporating bidirectional message passing, edge updates, and port-aware features, were compared with XGBoost and Random Forest using identical features. The best model was evaluated using GNNExplainer against documented typologies and Mondrian conformal prediction for uncertainty calibration. Results: IMPROVED achieved the strongest performance (AUPRC = 64.85%, Best F1 = 68.42%), exceeding BASE-Large by 30.65 AUPRC points and XGBoost (AUPRC = 38.46%) by 26.4 points. The proposed mechanisms contributed more than capacity scaling alone. GNNExplainer recovered documented typologies with higher fidelity than attention-weight and random baselines (mean Jaccard overlap: 21.4% vs. 1.0%, p < 0.001). Mondrian conformal prediction achieved coverage close to the 90% nominal target with an average prediction-set size near one. Conclusion: Explicit transaction topology modeling substantially improves AML detection, while faithful explanations and calibrated uncertainty support interpretable, human-in-the-loop compliance decision-making.
1. Introduction
Rule-based AML monitoring struggles with false positives and relational laundering patterns, while graph methods address transaction topology but leave explanation faithfulness and calibrated uncertainty underexamined. This study develops a graph-based framework to assess detection, explanations, and confidence together.
1. Introduction
- Rule-based monitoring often misclassifies legitimate behavior and generates false-positive alerts, while laundering patterns can depend on transaction structures across connected accounts.The rarity of illicit transactions and need for auditable decisions also make prediction performance alone insufficient.Source: §1 · p. 3
- Graph learning fits AML because accounts and transfers form a directed attributed graph, and laundering typologies include fan-out, cycle, and gather-scatter structures.These patterns are defined across connected accounts rather than by an isolated transaction.Source: §1 · p. 3
- Prior graph AML studies report gains from topology-aware architectures, but few evaluate explanation faithfulness against documented typologies or statistically calibrate prediction confidence.Reported approaches include reverse message passing, multi-edge aggregation, temporal models, and attention-based relational features [7] [9].Source: §1 · pp. 3–5
- The study addresses the literature gaps with a capacity-matched control, typology-based tests of GNNExplainer, and Mondrian class-conditional conformal prediction.Explanation evaluation uses precision, recall, and Jaccard overlap with paired statistical significance testing.Source: §1 · p. 5
2.1. Dataset Description
The study uses IBM's synthetic HI-Small AML benchmark, a severely imbalanced transaction dataset spanning accounts, institutions, currencies, and payment formats.
2.1. Dataset Description
- The synthetic HI-Small benchmark contains 5,078,345 transfers among 518,573 unique accounts, with 5,177 illicit transactions (0.10%).The account registry covers 20,053 institutions and 166,207 entities.Source: §2.1 · pp. 5–6
- Transactions span 15 currencies, 7 payment formats, and 18 days; 591,212 transfers (11.64%) are self-loops, predominantly Reinvestment-type transfers.Table 1 summarizes the dataset's statistical characteristics.Source: §2.1 · p. 6, Table 1 · p. 6
- The benchmark is publicly available and synthetic, with no real individuals or identifiable personal or institutional data, so ethical approval and informed consent were not required.The dataset includes ground-truth illicit transaction labels.Source: §2.1 · pp. 5–6
2.2. Graph Data Pipeline: Preprocessing, Feature Engineering, and Construction
The pipeline cleans transaction records, derives account and transfer features, and constructs a directed attributed graph for downstream modeling.
2.2.1. Data Preprocessing
- Preprocessing removes duplicate accounts, validates references, separates self-loops, and imputes missing aggregated values with zero.Removing self-loops reduces the transaction edge set from 5,078,345 to 4,487,133.Source: §2.2.1 · p. 7
2.2.2. Feature Engineering
- Account features aggregate incoming and outgoing transaction statistics, retain reinvestment behavior, and one-hot encode entity type; country is excluded because it has 149 unique values.Amount-based features are log-transformed; transaction features include log amount, cyclical hour, day index, payment format, and currency mismatch.Source: §2.2.2 · pp. 7–8
2.2.3. Graph Construction
- Graph construction maps accounts to node indices and transactions to directed edges, yielding 518,573 nodes and 4,487,133 edges.The graph stores node features, edge features, and binary illicit-transaction labels.Source: §2.2.3 · p. 8
2.3. Model Architecture and Training Configuration
The study compares three GATv2 architectures and adds bidirectional message passing, edge updates, and port features in IMPROVED. It also specifies leakage-aware feature standardization, loss functions, and training procedures.
2.3.1. Feature Standardization
- Edge-feature standardization uses statistics computed only from training edges to prevent validation and test information leakage.Node and edge features are z-score normalized; the standardization uses a numerical-stability constant of 10^-6.Source: §2.3.1 · p. 9
2.3.2. Graph Neural Network Architectures
- IMPROVED adds bidirectional message passing, edge updates, and port features to the capacity-matched BASE-Large architecture.BASE-Large uses three layers and hidden dimension 128; IMPROVED shares those dimensions, helping separate architectural additions from model scaling.Source: §2.3.2 · p. 11, §2.3.3 · p. 11
2.3.3. Edge-Update and Port Features (IMPROVED)
- Port features encode each transaction’s local rank among parallel edges at its source and destination, cueing fan-in/fan-out typologies.The bidirectional graph is constructed separately for each split to prevent temporal leakage.Source: §2.3.3 · p. 11
2.3.4. Loss Functions
- BASE and BASE-Large use class-weighted binary cross-entropy, while IMPROVED uses focal loss to emphasize hard, minority-class transactions.Focal loss uses class-balancing weights α_t ∈ {0.90, 0.10} and focusing parameter γ = 2.0.Source: §2.3.4 · pp. 11–12
2.3.5. Training Configuration
- All models use mini-batch neighbor sampling, and edge-classification representations concatenate source and destination embeddings with raw edge features.Optimization uses Adam and a plateau-based learning-rate scheduler monitoring validation AUPRC.Source: §2.3.5 · p. 12
2.4. Tabular Machine Learning Baselines
The tabular baselines receive the same raw transaction, source-account, and destination-account information as the GNNs, but without graph message passing. Their training addresses class imbalance and uses repeated-seed reporting.
2.4. Tabular Machine Learning Baselines
- XGBoost and Random Forest use source-node, destination-node, and edge features without message passing or neighborhood aggregation.The concatenated representation has 39 dimensions per transaction.Source: §2.4 · p. 13
- Class weighting addresses imbalance, while negative-class subsampling is applied only during training and evaluation retains the full class distribution.All positive instances are retained in training; validation and test sets are not subsampled.Source: §2.4 · pp. 13–14
- Each baseline is evaluated across five random seeds, with results reported as mean ± standard deviation.Table 4 summarizes the tabular-baseline configurations.Source: §2.4 · p. 14, Table 4 · p. 14
2.5. Model Explainability and Faithfulness Evaluation
The evaluation compares post-hoc edge explanations with attention and random baselines against labeled laundering-ring structure. It scores explanations within each target edge’s receptive field and reports ring coverage separately.
2.5. Model Explainability and Faithfulness Evaluation
- Each illicit test transaction in a labeled laundering ring is matched to its graph edge, and the ring’s remaining edges form the ground-truth explanation set.The annotated typologies include fan-out, cycle, and gather-scatter.Source: §2.5 · p. 15
- GNNExplainer and attention-weight explanations are compared with a random edge-selection baseline for each sampled transaction.GNNExplainer learns a soft edge mask by optimizing mutual information between the masked subgraph and the model prediction.Source: §2.5 · p. 15
- Explanations are restricted to the target edge’s K-hop receptive field, so faithfulness is scored against reachable ring edges and ring coverage is reported separately.A laundering ring may extend beyond the neighborhood.Source: §2.5 · p. 15
- The evaluation uses top-k precision, recall, and Jaccard overlap, with k set to the reachable ground-truth ring size.The paired Wilcoxon signed-rank test assesses differences in per-transaction Jaccard scores.Source: §2.5 · pp. 15–16
2.6. Uncertainty Quantification via Conformal Prediction
The procedure calibrates class-specific nonconformity thresholds on a held-out validation split, then forms prediction sets to represent confident or ambiguous transaction classifications.
2.6. Uncertainty Quantification via Conformal Prediction
- Mondrian conformal prediction calibrates a separate nonconformity threshold for each true class, supporting class-conditional coverage despite severe class imbalance.The procedure uses LAC nonconformity scores and thresholds estimated from the held-out validation set.Source: §2.6 · pp. 16–17
- At test time, each transaction receives an illicit, normal, or both-label prediction set, with two-label sets treated as ambiguous and flagged for human review.Empty sets are also possible in principle.Source: §2.6 · p. 17
- Calibration uses only the validation split, while coverage and average prediction-set size are measured on the untouched test split across four miscoverage levels.α = 0.10, corresponding to a 90% target coverage, is the primary reported level.Source: §2.6 · pp. 17–18
2.7. Data Splitting Strategy and Evaluation Metrics
The evaluation uses chronological, percentile-based splits to prevent temporal leakage and reports metrics suited to the dataset’s severe class imbalance. Model selection and final performance assessment use separate validation and test data.
2.7.1. Temporal Train/Validation/Test Splitting
- Percentile-based temporal boundaries address sharply non-uniform daily transaction volumes while avoiding the future-to-past leakage that random sampling could permit.The transaction volume declines steeply after the observation window’s tenth day.Source: §2.7.1 · p. 20
- The chronological split assigns transactions by the 70th- and 85th-percentile timestamps, yielding 2,590,591 training, 658,773 validation, and 1,237,769 test edges.The validation set is reserved for selection, scheduling, and conformal calibration; the test set remains untouched until final evaluation.Source: §2.7.1 · p. 20
2.7.2. Evaluation Metrics
- AUPRC is the primary model-selection criterion because it is more sensitive to illicit-class performance than AUROC under severe class imbalance.Accuracy is not used because only 0.10% of transactions are illicit.Source: §2.7.2 · p. 20
- Best F1 and its corresponding precision, recall, and threshold are selected by sweeping the decision threshold across the precision-recall curve.Metrics are computed on validation data for model selection and independently recomputed on the held-out test split for final reporting.Source: §2.7.2 · p. 21
3. Results
On the HI-Small benchmark, IMPROVED outperformed the other graph architectures and tabular baselines, while explanation and conformal evaluations assessed faithfulness and uncertainty. The results emphasize the gains from relational mechanisms and the practical behavior of calibrated prediction sets.
3.1. Experimental Setup
- The evaluation covers 1,237,769 held-out test transactions, including 2,094 illicit instances, from the temporally partitioned HI-Small benchmark.The comparison includes three GNN variants and two tabular baselines; the best model also undergoes explanation-faithfulness and conformal-uncertainty evaluation.Source: §3.1 · p. 22
3.2. Comparative Performance of Graph Neural Network Architectures
- 30.65 percentage points: IMPROVED’s test AUPRC rose from BASE-Large’s 34.20% to 64.85%, with 68.42% Best F1 versus 40.15% for BASE-Large.IMPROVED matches BASE-Large in depth and hidden dimensionality; the reported gain is associated with bidirectional message passing, edge updates, and port features.Source: Figure 3 · p. 23, Table 8 · p. 23, §3.2 · p. 24
3.3. Explanation Faithfulness Against Ground-Truth Laundering Typologies
- BIPARTITE samples had no reachable ground-truth edges within any model’s K-hop receptive field and were excluded from the typology table.The reported faithfulness scores therefore reflect reachable ground truth, not necessarily the entire laundering ring.Source: §3.3 · pp. 27–28
- GNNExplainer outperformed attention and random baselines in pairwise Jaccard comparisons for every model, with all three pairwise orderings statistically significant at p < 0.05.For the 2-hop FAN-OUT case, IMPROVED recovered two of three ground-truth edges, versus one of three for BASE and BASE-Large.Source: Figure 4 · p. 25, §3.3 · pp. 26, 28
3.4. Calibrated Uncertainty via Mondrian Conformal Prediction
- At α = 0.10, all three models achieved 89.6–90.7% overall coverage with average prediction-set size 0.999–1.000; illicit-class coverage met or exceeded target across the tested levels.At α = 0.05 and 0.01, empty sets vanished entirely.Source: Figure 5 · p. 29, Table 11 · p. 29, §3.4 · pp. 30–31
3.5. Comparison Against Tabular Machine Learning Baselines
- Across five models, test AUPRC was 64.85% for IMPROVED, 38.46% for XGBoost, and 17.97% for Random Forest on identical 39-dimensional features.IMPROVED exceeded XGBoost by 26.4 points in AUPRC but trailed it by 17.4 points in precision.Source: Figure 6 · p. 32, §3.5 · p. 33, Table 12 · p. 33
4. Discussion
On the synthetic HI-Small benchmark, IMPROVED combined the strongest detection results with more faithful typology explanations and near-nominal conformal coverage. The discussion relates these findings to prior AML work, the mechanisms and costs of the model, and limits on generalization and explanation scope.
4.1. Summary of Principal Findings
- 64.85% test AUPRC and 68.42% Best F1 for IMPROVED were the strongest results among evaluated models, with architectural additions contributing more than capacity scaling alone.The study also reports that IMPROVED outperformed tabular baselines and that its explanations recovered documented typologies with higher fidelity than attention-weight and random baselines; Mondrian prediction achieved near-nominal coverage with sets near size one.Source: §4.1 · p. 34
4.2. Comparison with Existing Studies on the Shared IBM AML Benchmark
- Table 13 compares studies using HI-Small or related AMLworld datasets; most comparator studies emphasize detection rather than explanation faithfulness or calibrated uncertainty.The discussion also places the present model's architecture and results alongside this prior work, noting that its detection scores fall within a comparable range to related architectures.Source: §4.2 · pp. 34–35, Table 13 · p. 35, SALT-GNN · p. 36
SALT-GNN
- In IMPROVED, the introduced bidirectional edge-update and port-aware mechanisms yielded gains of 25 to 31 percentage points; most listed studies, including Multi-GNN and its multi-edge extension [7], focused on detection performance.Blanuša et al. [9] describe an interpretable smurfing-score approach that is explainable by construction, but it is unsupervised and was not evaluated against ground-truth laundering typologies using post hoc attribution methods or paired with formal uncertainty quantification.Source: SALT-GNN · p. 36
4.3. Methodological and Practical Implications
- Bidirectional propagation carries information both with fund flow and in the direction of suspicion; edge updates refine transaction representations using neighboring context, while port features encode local transaction rank.These mechanisms are discussed as relevant to fan-out and gather-scatter patterns; Focal Loss down-weights the many easily classified licit transactions and focuses learning on hard minority examples.Source: §4.3 · p. 37
4.4. Interpretability and Its Role in Regulatory and Operational Trust
- GNNExplainer outperformed raw attention weights across architectures, while STACK's 4.09% Jaccard for IMPROVED was mainly constrained by 24.41% mean receptive-field coverage.The discussion recommends reporting receptive-field coverage alongside explanation metrics to distinguish localized structural limits from explainer quality.Source: §4.4 · pp. 37–38
4.5. Uncertainty Quantification for Human-in-the-Loop Deployment
- At α = 0.10, all three models achieved 89.6–90.7% overall coverage with average prediction-set sizes of 0.999–1.000; IMPROVED produced more confident illicit singletons than the other models.The discussion presents this as a way to resolve most transactions to confident singleton decisions while routing a residual fraction for review.Source: §4.5 · pp. 38–39
4.6. Computational Environment and Training Efficiency
- IMPROVED took 332.45 minutes to train for 100 epochs, versus 201.20 minutes for BASE-Large and 98.33 minutes for BASE.The discussion attributes the longer IMPROVED training time to bidirectional message passing, edge updates, and port-feature computation, and flags scaling to larger datasets or real-world volumes as a consideration.Source: §4.6 · pp. 39–40, Table 14 · p. 39
4.7. Limitations and 4.8. Future Work
- Results are limited to a synthetic HI-Small benchmark; generalization to proprietary data, cross-border signals from the excluded country feature, and precision–recall trade-offs remain open considerations.Future work includes evaluating larger dataset variants and institutional data, encoding country attributes, adding explicit temporal dynamics, and tailoring receptive fields to typology depth.Source: §4.7 · p. 40, §4.8 · pp. 40–41
5. Conclusion
The study concludes that IMPROVED's graph architecture produced the strongest AML detection results, while explanations and conformal prediction added support for trustworthy human-in-the-loop use. These findings are scoped to the evaluated transaction network, and broader validation remains relevant because the dataset is synthetic.
5. Conclusion
- Explicitly modeling directed, multi-edge transaction structure was a more effective lever for detection performance than scaling model size alone.IMPROVED achieved the strongest performance among evaluated models and tabular baselines, with bidirectional message passing, edge updates, and port-aware features.Source: §5 · p. 41
- GNNExplainer recovered documented typologies with higher fidelity than attention and random baselines, while Mondrian prediction provided near-nominal coverage and sets near size one.The conclusion frames these properties as support for auditable rationales and separation of confident decisions from cases warranting human review.Source: §5 · pp. 41–42
- The evaluation used the synthetic IBM HI-Small dataset and did not use proprietary, institutional, or human-subject data.The authors note that further validation would help establish how results transfer to real-world financial data.Source: §5 · pp. 41–42